Performance & Cost, Shown
Every performance and cost figure we publish is either a measurement you can reproduce or a calculation from published prices. This page shows the work. Numbers below were measured on July 24, 2026.
Measured performance
Benchmarked on the reference GPU host
(Intel i9-14900KF, RTX 4070 Ti SUPER 16GB) running the default
Tier 1 model qwen3.5:9b via Ollama,
warm (model resident). Reproduce with
python manage.py test_ollama_models --benchmark.
| Task | Tokens out | Throughput | Wall time |
|---|---|---|---|
| Classification | 7 | 85.7 tok/s | ~0.3 s |
| Business analytics | 1183 | 85.4 tok/s | ~22 s |
| Code generation | 1382 | 95.5 tok/s | ~14 s |
| Strategic planning | 2293 | 88.0 tok/s | ~26 s |
A warm classification (a handful of output tokens) returns in ~300 ms; a short ~60-token answer in ~890 ms. Average generation throughput across task types is ~88.6 tok/s.
gpt-oss:120b-cloud)
whose latency is network-bound and not represented here. Escalated (Tier 2/3)
requests to a frontier model are slower than local generation. We publish only
what we measured.
The cost model
MythosLab routes by stakes, not raw complexity. The Tier 1 workhorse handles the bulk of work at a flat or zero marginal cost; only the hard or accuracy-critical minority reaches a per-token frontier model. Published list prices (per million tokens):
| Tier | Model | Input | Output | Share of traffic |
|---|---|---|---|---|
| 1 workhorse | qwen3.5:9b / gpt-oss:120b-cloud |
$0 local / flat | $0 local / flat | 85-90% |
| 2 general | claude-sonnet-5 |
$3.00 | $15.00 | 5-8% |
| 3 critical | claude-opus-5 |
$5.00 | $25.00 | 2-5% |
Worked example. Assume an average request of ~500 input + ~1,000 output tokens, and the tier mix above (88% / 7% / 5%). Every escalated request costs $0.0165 (Sonnet 5) or $0.0275 (Opus 4.8). Two monthly volumes:
| Monthly requests | Per-token escalation | On your own GPU | On a GPU-less VPS |
|---|---|---|---|
| 5,000 | ~$13 | ~$13/mo | ~$33/mo |
| 10,000 | ~$25 | ~$25/mo | ~$45/mo |
The VPS column adds a flat ~$20/mo Ollama Cloud subscription for Tier 1 (a GPU-less box cannot run the local model); the GPU column runs Tier 1 for free on hardware you already own. Embeddings always run locally on CPU at no cost.
What it replaces
Same workload, no tiering. Running those same 10,000 requests entirely on a frontier model instead of tiering costs ~$165/mo (all Sonnet 5) to ~$275/mo (all Opus 4.8) in API charges alone. Tiering cuts that by 60-85% -- the same math, just keeping the routine work off the meter.
Versus per-seat enterprise AI. Hosted enterprise AI assistants are priced per seat: Microsoft 365 Copilot at $30/user/mo on top of a qualifying license (all-in $69-90/seat), ChatGPT Enterprise reported around $45-75/seat with a 150-seat minimum, and enterprise search tools like Glean at $40+/seat plus a platform fee. For a 20-person business that is $600-1,800/month, scaling with headcount. MythosLab's compute cost does not scale per seat -- the figures above cover the whole team.
Sources & assumptions
- Local benchmarks:
test_ollama_models --benchmarkon the reference host, July 24, 2026. Your throughput varies with hardware and prompt shape. - Anthropic list prices are the model rates MythosLab charges cost against in
core/inference/anthropic_client.py. - Tier shares (85-90% / 5-8% / 2-5%) are routing targets; your real mix depends on your workload and is visible on the in-app analytics dashboard.
- Enterprise-AI per-seat pricing reflects mid-2026 published/reported rates and is illustrative; exact quotes vary by contract and existing licensing.